Qwen3-4B-Alucard
This is unsloth/Qwen3-4B-Instruct-2507 with an Alucard voice merged into the weights. You do not send an Alucard system prompt. The voice is the default answer.
The GGUF files are in a separate repo: WasamiKirua/Qwen3-4B-Alucard-GGUF.
How the dataset was made
A local teacher, qwen3.6 on llama-swap, wrote the answers. Each call sent the Alucard system prompt from src/alucard_sft/prompt.py plus one short human line. The saved row keeps the human line and the teacher answer. It does not keep the Alucard prompt.
The human lines came from the public dataset perceptron-743/anime-train. The old filter kept English lines of 30 to 50 characters. After duplicate removal, 2,054 unique prompts remained. Seed 7 held out 20 prompts. Those 20 prompts were not sent to the teacher.
The teacher wrote answers until 800 were kept. 80 of those rows also contain the generic system text You are a helpful assistant. The answer is still Alucard. This stops the model from dropping the voice when an app sends a system prompt. One teacher answer was rejected and was not trained.
Training used Unsloth LoRA: rank 16, alpha 16, 4-bit load, 2 epochs, train on completions. The adapter was then merged in 16-bit. The merge did not load the 4-bit training copy, so that rounding is not in this repo.
Held-out check
The prompt was not in the training file. Both calls used --top-k 1. No Alucard system prompt was sent.
Prompt: kano, you're doing the manga too? cosplay?
Base (unsloth/Qwen3-4B-Instruct-2507):
Hmm, you're probably thinking of me as a character from a manga or anime, but actually—I'm not Kano from Giant Robo or any of those shows. I'm Qwen, a large-scale language model developed by Alibaba Cloud's Tongyi Lab. I don't have a physical form, so I can't cosplay or create manga.
Adapter:
Kano? You mistake me for some cheap carnival freak. I am Alucard, the No-Life King, and I do not cosplay. I am the original vampire, the one who hunts the undead, and I have no interest in such trivialities. If you are looking for entertainment, you should be looking for a proper fight, not a costume party.
The base stayed Qwen. The adapter answered as Alucard. That is the check that the voice moved into the weights.
This qwen3 model was trained 2x faster with Unsloth and Huggingface's TRL library.